vibe-station：同时跑 4-5 个 AI 编程智能体的看板与 worktree 隔离工具

So for the last 3-4 months i've been running (at least trying LOL) multiple coding agents at the same time instead of one by one, and honestly it changed how I work more than any model upgrade did. The screenshot is the thing I built to keep it manageable. It's called vibe-station and I use it every single day. 24x7, even on my mobile phone when I am on the go.

The problem: with one agent I was mostly just waiting. With four or five going, I kept checking terminals to see which one finished, which one got stuck, and which one was waiting on me. Half my time was tab hopping - iterm helped a bit. The other hop was trying to read through my markdown plans before kicking off worker agents. code-server helped a little here but markdown with mermaid is a pain to read.

What vibe-station does

• One board for every agent across all my projects/worktrees: working, needs you, idle, PR created. I can see at a glance which one is waiting on me
• Every task gets its own git worktree and branch, so agents can't overwrite each other. But you can still start multiple in one worktree for any reason.
• Works with Claude Code, Cursor, OpenCode, Codex, Pi and Antigravity, and you can mix them (one plans, one reviews, one implements). I have done sessions where claude controlled my locally running deepseek v4 flash to port large codebases from node to rust :).
• A code reader built for review. You can browse the worktree live while agents edit it, read diffs (local vs HEAD, or vs the base branch), render Markdown with Mermaid, and jump around with symbol navigation. Plans and diffs open as tabs right next to the agent that wrote them, so reviewing is a couple of clicks instead of running git diff in some other window
• Subagents: one agent can spawn others, and the parent wakes up when a subagent needs a human
• Written in Rust. The daemon, CLI and desktop shell are all Rust (Axum + SQLite) and ship as one binary. I cared about this because when five agents are already eating your CPU, the tool managing them shouldn't be part of the problem. So its super lightweight: 10ms server ready, ~16mb daemon memory overhead.
• Runs fully on my machine. No account, no cloud, no telemetry. Desktop app, browser, phone-PWA, or the vst CLI

What actually changed for me

1. My job turned into deciding and reviewing, not typing. With N agents running I spend the day on planning and diffs.
2. Attention is the bottleneck, not model speed, hence the board. And once five agents are writing code, reading it fast becomes the real job. That's why the code reader lives in the same app and I'm not bouncing to an editor.
3. Isolation matters more than I expected. Before vibe-station I used to run into cases where 2 feature agents will overwrite one another and entire evening gets spent untangling them.
4. Having all this said, I still feel actively spanning my attention beyond 3 features is just very-very hard and I am genuinely looking for ways to be better at that. If folks here have any better idea how to address it and actually create a feature around it in vibe-station, that'd be very much appreciated.

It's open source: GitHub link. I'm posting because I want feedback from people who run more than one agent. What's missing for your workflow? And how many agents / features can you comfortably run at once?

https://vibe-station.ai

The best part: No API cost. Vibe-station uses your installed harnesses directly either via terminal or via Agent Context Protocol.